On Fri, Dec 09, 2011 at 06:02:21PM +0000, Ramana Radhakrishnan wrote:
> On 8 December 2011 21:06, Richard Henderson <r...@redhat.com> wrote:
> > ---
> > gcc/config/arm/arm-protos.h | 3 +
> > gcc/config/arm/arm.c | 527
> > ++++++++++++++++++++++++++++++++-
> > gcc/config/arm/neon.md | 59 ++++
> > gcc/config/arm/vec-common.md | 26 ++
> > gcc/testsuite/lib/target-supports.exp | 9 +-
> > 5 files changed, 620 insertions(+), 4 deletions(-)
>
> I haven't been following the vector permute work in great detail and
> I must say I haven't read this patch series in great detail yet.
>
> For Neon a further optimization to consider might be to use the vext
> instruction which could achieve permute masks that are monotonically
> increasing constants ? While I expect the latency for a vext or vtbl
> instruction to be about the same (your mileage might vary depending on
> the core), using vext gives us the freedom of not needing a register
> for the permute mask -
>
> a = vec_shuffle (b, c, mask) where mask is { n + 7, n + 6, n + 5, n +
> 4, n + 3, n + 2, n + 1, n } could just be vext.8 A, B, C, #n
>
> If the mask being provided is a reverse of the mask above, it's
> probably not worth it.
>
>
> Additionally , can we also detect rotate rights ? unless ofcourse
> there's a different interface -
>
> a = vec_shuffle (vec, {0, 7, 6, 5, 4, 3, 2, 1}) => vext.8 a, vec, vec, #1
>
>
> Masks doing rotate lefts are probably not worth it in this
Richard and I were discussing this last night on IRC, and it is certainly
possible. Somebody would just have to write a predicate to recognize the
case. We do wonder how frequently it will occur, and whether people doing this
would just use the whole vector shift instead of shuffle.
--
Michael Meissner, IBM
5 Technology Place Drive, M/S 2757, Westford, MA 01886-3141, USA
meiss...@linux.vnet.ibm.com fax +1 (978) 399-6899
